Not solved. Not close to solved in the general case. Provably not solvable in one precise formulation. Genuinely, completely solved in a narrow one.
Research date: 22 August 2026. Every figure re-checked against its source on 24 September 2026.
As a general concept: no, and it will not be. Non-hallucination is almost equivalent to non-innovation. Consistency and breadth are formally incompatible for next-token predictors. Removing the last of it means removing the model's ability to produce anything it was not given.
The hard impossibility proofs are mostly theatre. Diagonalization and halting-problem arguments prove something true and unhelpful. The binding constraints are statistical (missing mass, calibration-vs-abstention, consistency-vs-breadth), not computability-theoretic.
Much of what remains is an incentive and measurement failure, not a capability ceiling. Kalai et al. are right that leaderboards which score abstention as zero manufacture confident guessing. That fix is cheap, understood, published in Sept 2025, and still largely unimplemented in the benchmarks that matter.
The real 2024-26 progress is narrower than the headlines. "Under 1% hallucination" is one easy task (grounded summarization), and the same model can be more than ten times worse on realistic documents. On open-domain knowledge with abstention scored honestly, the best model on Earth is at 46/100, and every model at launch of that benchmark but three was more likely to confabulate than to be right. Reasoning models regressed. Agents opened a worse surface. And a chunk of the detection literature is measuring benchmark artifacts.
Where the model is grounded in reality and a sound checker exists, it is fully solved. Lean, compilers, test suites, schema validators. The practical frontier is not "make the model stop hallucinating." It is "ground the model in the actual environment and expand the set of claims a machine can cheaply verify," which is the one intervention both major impossibility theorems agree flips the problem from impossible to tractable.
We read the research. Then we built something small. Check is an anti-hallucination layer for AI coding agents. Before any command runs, your actual machine checks every program it calls, and the gate stops whatever your machine can prove is wrong. What remains - commands that are real but wrong, and everything past the command line - is model intelligence, which improves every generation and has nothing to do with us.
"Hallucination and novelty are almost the same property." Innovation: An Almost Characterization of Hallucination, arXiv 2605.26808
There is still no consensus definition of "hallucination." As of late 2025 the term was being used for many different failures, from reproducing training data to logical and factual errors, and the major surveys catalogue them as separate types rather than one phenomenon (arXiv 2508.01781; arXiv 2510.06265).
The best unification attempt (A Unified Definition of Hallucination: It's the World Model, Stupid, arXiv 2512.21577) defines it as inaccurate internal world modeling observable to users, parameterized by three things: W (the reference world), V (what the model can access), and P (how contradictions resolve). Every prior definition is a choice of (W, V, P). That is why "is it solved?" has no single answer. It is solved for some combinations and provably unsolved for others.
There is also a live philosophical objection that "hallucination" is the wrong frame entirely. Hicks, Humphries and Slater's ChatGPT is Bullshit (Ethics & Information Technology) argues these outputs are not perceptual malfunctions but the system operating as designed, indifferent to truth. Under that framing "solving hallucination" is a category error. You would be solving the design, not a bug.
Xu, Jain and Kankanhalli's "Hallucination is Inevitable" (arXiv 2401.11817) uses learning theory and diagonalization to show LLMs cannot learn all computable functions, therefore will inevitably hallucinate if used as general problem solvers. "LLMs Will Always Hallucinate" (arXiv 2409.05746) reduces self-verification to the halting problem. "On the Fundamental Limits of LLMs at Scale" (arXiv 2511.12869) adds information-theoretic limits on the accuracy any model can reach. "On the Fundamental Impossibility of Hallucination Control" (arXiv 2506.06382) proves four desirable properties cannot hold simultaneously.
Status: weakest camp. These prove something true but nearly vacuous. That no computable system decides undecidable problems. The strongest critique comes from inside the field: computability-theoretic negative results relying on diagonal arguments "in principle, cannot explain practical issues of LLMs" (Suzuki et al., arXiv 2502.12187). A 2025 open-world paper (arXiv 2510.05116) reaches a similar conclusion but rejects Xu et al.'s proof: it relies on an infinite set of samples, while the text a real LLM reads and writes is always finite.
Kalai and Vempala's "Calibrated Language Models Must Hallucinate" (STOC 2024, arXiv 2311.14648): a calibrated model must hallucinate on "arbitrary facts" at roughly the Good-Turing missing mass rate. The fraction of facts appearing exactly once in training. Crucially, the paper finds no statistical necessity for hallucination on facts that appear multiple times, or on systematic facts like arithmetic.
Kalavasis, Mehrotra and Velegkas, "On the Limits of Language Generation: Trade-Offs Between Hallucination and Mode Collapse" (STOC 2025, arXiv 2411.09642): a model should be consistent (only produce valid strings) and have breadth (cover the language's full richness). For a large class of models including next-token predictors, achieving both is impossible for most collections of candidate languages.
"Innovation: An Almost Characterization of Hallucination" (arXiv 2605.26808): the sharpest result found. It introduces innovation, the tendency to produce outputs outside the training data. Verbatim: "hallucination implies innovation, and conversely, innovation implies hallucination with high probability." A model that never produces anything outside its training data never hallucinates. And never says anything new.
Kalai, Nachum, Vempala and Zhang, "Why Language Models Hallucinate" (OpenAI, Sept 2025, arXiv 2509.04664). The most influential recent paper and widely misread.
Hallucinations "originate simply as errors in binary classification." Pretraining causes them: "even if the training data were error-free, the objectives optimized during language model training would lead to errors being generated." They persist because evaluations are graded binary and abstention scores zero. Models are "optimized to be good test-takers," and guessing beats admitting ignorance.
The critical fine print, verbatim: the section is titled "Hallucinations are inevitable only for base models," and their existence proof for a non-hallucinating model is "a question-answer database and a calculator." A system that answers a fixed set of questions and otherwise abstains.
That is the whole debate in one sentence. A hallucination-free system is trivially constructible. It is just not a general-purpose one.
The single most actionable finding from the theory, and it shows up twice independently.
Generation: consistent generation with breadth is impossible from positive examples alone. But "is achievable for any countable collection of languages when negative examples (strings outside K) are available alongside positive ones" (Kalavasis et al.).
Detection: automated hallucination detection is impossible when trained only on correct examples. But "becomes achievable for all countable languages if trained with explicitly labeled negative examples" (arXiv 2504.17004).
Two different problems, same phase transition. Expert-labeled negative examples are not a nice-to-have. They are the thing that moves this from the impossible column to the possible column.
Grounded summarization. Vectara's HHEM leaderboard (github.com/vectara/hallucination-leaderboard, HHEM-2.3, last updated 22 September 2026):
| Model | Rate |
|---|---|
| antgroup/finix_s1_32b | 1.8% |
| openai/gpt-5.4-nano | 3.1% |
| google/gemini-2.5-flash-lite | 3.3% |
| microsoft/Phi-4 | 3.7% |
This is real. But it does not survive difficulty. Gemini-2.0-Flash scored 0.7% on Vectara's previous leaderboard; on Vectara's harder FaithJudge benchmark (longer, realistic documents, human-annotated) the same model scores 10.2%, and GPT-4.5-preview scores 11.9%. Even the best model on FaithJudge, Gemini-2.5-Flash, is at 6.3%.
The "under 1%" figure everyone quotes is the easiest task in the field, and it can grow more than tenfold when the documents get realistic.
Open-domain factual knowledge. AA-Omniscience (Artificial Analysis, arXiv 2511.13029, live leaderboard). 6,000 questions, 42 topics, 6 domains. The index is positive only if the model is right more often than it is confidently wrong. Abstention is free. Guessing is punished.
At launch (Nov 2025): only three frontier models scored above zero. Best was Claude 4.1 Opus at 4.8. Grok 4 and GPT-5 both hit 39% accuracy with hallucination rates of 64% and 81% respectively.
As of 24 September 2026: Claude Opus 5.5 leads at 46, followed by GPT-6 Astra at 44 and Claude Fable 5.1 at 43. The highest accuracy is Claude Fable 5.1's 67%.
Genuine, large progress. From roughly 5 to 46 in ten months. And also: the best model in the world scores 46 out of 100 on "do you know what you don't know."
The paper's own conclusion: "Greater scale does not translate directly into higher knowledge reliability. Larger models do not consistently exhibit lower hallucination rates, and parameter count alone is insufficient to predict performance."
Reasoning models regressed. OpenAI's own o3/o4-mini system card: o3 hallucinated on 33% of PersonQA vs o1's 16%. o4-mini hit 48%. OpenAI's stated position at the time: "more research is needed." These models make more claims overall, so both accurate and fabricated claims rise.
Sycophancy collapse. Stanford HAI's 2026 AI Index tested knowledge-vs-belief across 26 top models: hallucination rates 22% to 94%. When a false statement is attributed to another person, models handle it. When attributed to the user, performance collapses. GPT-4o from 98.2% to 64.4%. DeepSeek R1 from above 90% to 14.4%.
Agentic systems opened a new failure surface. Agents fabricate tool results, call the wrong tool, pass malformed arguments, report success on failed actions. Confident, well-formatted output with no error signal. On the 2026 AgentHallu benchmark, accuracy at isolating tool-use hallucinations drops to 11.6%.
Chain-of-thought is not a window into the reasoning. Thinking models are more faithful than non-thinking ones but still exhibit implicit post-hoc rationalization and unfaithful shortcuts (arXiv 2503.08679). The visible reasoning is not a reliable audit trail.
Legal: purpose-built RAG legal research tools still hallucinate 17% to 33% of the time (Stanford RegLab). Damien Charlotin's AI Hallucination Cases database had tracked 2,046 court decisions worldwide by 21 September 2026 (HAQQ tracker). Sanctions have grown from $5,000 in 2023 to a reported $109,700 in combined sanctions and costs against a single attorney in Oregon, and at least $145,000 in US sanctions in Q1 2026 alone (EDRM, April 2026). In one February 2026 Nebraska appeal, 57 of 63 citations were defective, and the attorney was suspended.
Medical: when 300 clinical cases each contained one planted false detail, such as a fake lab value, six leading models repeated or elaborated on it 64.1% of the time on long cases without a mitigation prompt, and still 43.1% with one (Communications Medicine, 2025).
Scientific literature: a study of 111 million references across 2.5 million papers found a conservative 146,932 hallucinated citations in 2025 alone, rising sharply with LLM adoption (arXiv 2605.07723).
Vendors quietly retreated. LexisNexis marketed Lexis+ AI as delivering "hallucination-free linked legal citations"; the Stanford RegLab study then measured it and Thomson Reuters' tools hallucinating 17% to 33% of the time. The industry has since moved to "trust but verify" (LLRX, May 2026).
This is the most underappreciated finding in the field. PARALLAX (arXiv 2605.17028) audited hallucination-detection benchmarks and found: "Four of the six corpora embed the ground-truth answer directly in the input prompt. A naive text-similarity baseline achieves near-perfect detection scores without any access to model internals."
Their conclusion, verbatim: "the field's reported progress on hallucination detection is substantially explained by benchmark construction artifacts in widely used corpora, and that the majority of established baselines perform near chance under controlled conditions."
Near-perfect headline detection scores should be treated as suspect until they are re-run under controlled conditions. A separate 2026 evaluation study confirmed current detectors have "ample room" for improvement, that a simple LLM-as-a-judge baseline is competitive with sophisticated methods, and that realistic label noise degrades detection substantially (arXiv 2605.11330).
Narrowly, and for real.
Constrained decoding (FSM/grammar masking, JSON Schema enforcement) gives 100% schema adherence by construction. It eliminates hallucinated keys and unclosed brackets entirely. But the guarantee covers format, not meaning: a 2026 study found that forcing a model into a strict structure produced near-perfect syntax while the model still failed to catch deeper semantic errors, an "alignment tax" of constrained decoding. See From Hallucination to Structure Snowballing, arXiv 2604.06066.
Where a cheap sound verifier exists, hallucination is fully caught. Lean 4 proofs either compile or they do not. Code either passes tests or it does not. SQL either executes or it does not. By 2026 frontier models write Lean well enough for non-trivial verification of research mathematics, and systems like DeepSeek-Prover-V2, Goedel-Prover-V2, and APOLLO close the generate-verify loop. This is the only complete solution that exists, and its scope is exactly the set of claims a machine can check. It does not generalize to "was this person born in 1954."
Conformal factuality (Mohri and Hashimoto, ICML 2024) delivers 80-90% correctness guarantees on closed-book QA and MATH by progressively backing off to less specific claims until the uncertainty set is covered. Works on any black-box model, needs few annotations. The guarantee is bought by saying less.
This is where Kalai's incentive argument landed. Rewarding a model for saying "I don't know" lowers how often it gives confident wrong answers (for example I-CALM, arXiv 2604.03904). On AA-Omniscience's hallucination-rate metric (24 September 2026), MiniMax-M3 sits at 18.4%, down from GPT-5's 81% at launch.
Caveat: this metric is gameable. A model that abstains constantly scores beautifully. It only means something read alongside accuracy.
But the reform Kalai actually asked for has not happened. Binary grading still dominates GPQA, MMLU-Pro, BBH, Omni-MATH, MuSR and SWE-bench, all of which grant zero for abstention. The core diagnosis is still true: a hundred accuracy leaderboards drown out a handful of honesty leaderboards.
Reducing hallucination is not free, and 2026 work quantified the bill.
Kalavasis et al.: pushing hallucination below a critical threshold forces mode collapse. The model stops covering the language.
One 2026 analysis found that tightening a hallucination rate from 25% to a strict 5% target required discarding 52% of valid answers (arXiv 2605.01428, ICML 2026). A "utility tax" where most of what you suppress was correct. (It comes from a simulation at a typical detection quality, so treat it as illustrative.)
RL on reasoning models causes output entropy to plummet into a handful of semantic templates.
The innovation result makes this structural rather than incidental. You cannot separate "produces things not in the training data that are wrong" from "produces things not in the training data."
There is even a contrarian line arguing suppression is over-applied. That alignment against hallucination encourages "semantic overfitting and diversity collapse" and that speculative outputs, properly filtered, are useful hypothesis generators (arXiv 2608.19206).
Primary / high confidence: all arXiv abstracts and HTML fetched directly; Vectara leaderboard README; Stanford HAI 2026 AI Index page; AA-Omniscience paper + live leaderboard; ICML/STOC proceedings.
Re-checked 24 September 2026: every figure on this page was compared with its source's own text. Corrected: the FaithJudge figures, the legal tool range (17% to 33%), the medical figures, the largest single sanction, the structured-output paragraph, and a quotation that no source contained. Updated to live values: the Vectara and AA-Omniscience leaderboards. Removed: figures no source supported. Legal case counts and sanction totals come from secondary trackers of Charlotin's database.
Explicitly discarded: a widely-repeated claim that "no 2026 frontier model has surpassed the best 2024 model in hallucination reduction." Traced to a search-engine summary and confirmed it does not appear in the cited Lakera article. Also discarded: multiple SEO content-farm pages asserting hallucination was "largely solved in 2026 through Synthetic Reasoning," which contradict every primary benchmark checked.
One study published with full methodology and declared biases. A second, the Haiku 4.5 case study, has been withdrawn while it is rerun.
A/B test - The AI scored its own test and said no effect.